Multimodal Deep Learning Information Fusion Diagnosis Method For Rotor Faults

As the key equipment in petroleum and the petrochemical industry,rotating equipment operating under severe condition is prone to various failures.The rotor system is the core part of the rotating equipment.Due to the complex structure,numerous components,frequent coupling failures,numerous and complex vibration of rotor system,it's difficult to cover the entire equipment and all types of faults by only using vibration monitoring.Infrared monitoring can simultaneously measure the temperature of multiple components,which is suitable for coupled fault diagnosis.Therefore,in this paper,a fault diagnosis method based on infrared images is introduced for rotor system.The study on the region of interest(ROI)extraction of infrared image,feature learning and fusion of infrared image and vibration signal are conducted.(1)The feature learning of infrared image and vibration signal based on convolutional neural network(CNN)is studied.Automatic extraction of infrared and vibration characteristics can be achieved by establishing CNN models which are suitable for infrared image and vibration information feature learning.Compared with the methods using artificial features,the accuracy of the proposed method has increased by 12.5 and 3.33 percentage points.Then,KPCA is conducted for reducing feature dimension and visualization,which directly proves that the CNN-extracted feature yields a better clustering distribution.(2)A ROI method based on infrared saliency detection and threshold optimization is proposed to solve the problems of intensity concentration,low contrast ratio,background interference on infrared image.The region contrast detection method is used to effectively remove the interference background in the image.The image segmentation threshold is iteratively optimized according to the diagnosis results of the random forest,the optimal extraction of ROI can be realized.Compared with the method using sensitive area selected by artificial frame,the accuracy of this method can increase by 7.5 percentage points?(3)Multimodal deep learning information fusion diagnosis method is proposed for addressing the problems of information fusion and fault diagnosis based on infrared and vibration.Feature learning and information fusion are combined into one model,and increase the correlation within the data by supervised-trained learning and reversing the weights of feature learning and fusion network.Compared with other decision level and feature level fusion diagnostic methods,it is shown that the fault recognition rate is higher by using this method.